Exploring Transformer Models and Domain Adaptation for Detecting Opinion Spam in Reviews

As online reviews play a crucial role in purchasing decisions, businesses are increasingly incentivized to generate positive reviews, sometimes resorting to fake reviews or opinion spam. Detecting opinion spam requires well-trained models, but obtaining annotated training data in the same domain (e.g., hotels) can be challenging. Transfer learning addresses this by leveraging training data from a similar domain (e.g., restaurants). This paper examines three popular transformer models—BERT, RoBERTa, and DistilBERT—to evaluate how training data from different domains, including imbalanced datasets, impacts Transformer model performance. Notably, our evaluation of hotel opinion spam detection achieved an AUC of 0.927 using RoBERTa trained on YelpChi restaurant data.